Intel is successfully experimenting with testing AI workloads on silicon photonics and bringing optical neural networks to life. The company’s scientists made the announcement through a paper novel, sharing the details of techniques that need to be used in order to achieve this.
The development comes after previous work’s proof of function about the success of a photonic circuit. Known as Mach-Zender interferometer or MZI commonly, this component can be configured in a way that it performs like a two-by-two matrix which emits the equivalent to relational phases quantitively between two light beams. When the smaller matrixes are arranged in a triangular mesh (making a larger matrix), they automatically create a circuit based on the principle of the matrix-vector multiplication, one of the fundamentals of deep learning.
“As in any manufacturing process, there are imperfections, which means that there will be small variations within and across chips, and these will affect the accuracy of computations,” wrote Intel AI products group senior director Casimir Wierzynski. “If ONNs is to become a viable piece of the AI hardware ecosystem, they will need to scale up to larger circuits and industrial manufacturing techniques. Our results suggest that choosing the right architecture in advance can greatly increase the probability that the resulting circuits will achieve their desired performance even in the face of manufacturing variations.”
Intel’s team tested with two architectures while building its AI experiment using MZIs — GridNet and FFTNet. The two were trained in simulation deep learning task of handwritten digit recognition (MNIST), benchmarked was such tasks. Here are the observations from the experiment.
|Grid like pattern||Butterfly like pattern|
|98% accuracy||95% accuracy|
|Drop in performance after adding the element of artificial noise. The performance fell below 50%.||No drop in performance. FFTNet is more robust than GridNet.|
Intel states that these observations imply that chips would need to to be modified/optimized after manufacturing — saving time and labor costs.